Skip to content

security-privacy.md

Latest commit

 

History

History
127 lines (97 loc) · 4.38 KB

security-privacy.md

File metadata and controls

127 lines (97 loc) · 4.38 KB

Security and Privacy

Design Principles

AIdeator is local-first by default. The privacy mode system ensures that data egress is explicit, auditable, and user-controlled.

Privacy Modes

local-only (default)

Guarantees:

  • Zero outbound HTTP calls during run execution
  • Idea title, description, target user, and context never leave the machine
  • Search signals are skipped entirely
  • LLM must be a local provider (e.g., Ollama)

Caveats:

  • Logs may include run IDs, timestamps, and status metadata
  • If a cloud LLM provider is configured but mode is local-only, search is still blocked but LLM calls depend on provider configuration
  • Avoid placing secrets in idea content fields

hybrid

Guarantees:

  • Search queries are truncated to 10 keywords maximum — no full idea text is sent
  • Only the search query payload leaves the machine (no description, context, or target user)
  • LLM outbound follows provider configuration

What external providers see:

  • A keyword-only search query like "AI code review developer tool SaaS"
  • LLM prompts (if a cloud LLM is configured)

cloud-enabled

Guarantees:

  • Full idea context may be sent to external search and LLM providers
  • Run payloads should be treated as potentially externalized data

What external providers see:

  • Full search queries including title + description
  • Full LLM prompts with idea context

Data Storage

Data Location Format
Ideas, runs, reports APP_DB_URL (default: SQLite file) In-memory dict or SQLite
Markdown reports APP_DOCS_DIR (default: ~/.local/share/aideator/docs) .md files
Config file ~/.config/aideator/config.toml TOML
Logs stdout/stderr Text or JSON

No data is stored on remote servers. All persistence is local to the configured paths.

API Keys

Storage

  • API keys are loaded from environment variables or the config file
  • Keys in config file are stored in plain text — use file permissions to restrict access
  • The config show command redacts secrets in output

Rotation

  • Keys can be rotated by updating the environment variable or config file
  • No session caching — new keys take effect on next server restart

Per-provider keys

Provider Environment Variable Required?
Tavily TAVILY_API_KEY Only if SEARCH_PROVIDER=tavily
Exa EXA_API_KEY Only if SEARCH_PROVIDER=exa
OpenAI LLM_API_KEY Only if LLM_PROVIDER=openai-compatible
Anthropic LLM_API_KEY Only if LLM_PROVIDER=anthropic-compatible
Mistral LLM_API_KEY Only if LLM_PROVIDER=mistral-compatible
Ollama — No key needed (local)
Built-in search — No key needed

Telemetry

  • No remote telemetry is configured by default
  • Logs go to stdout/stderr only
  • No analytics, tracking, or usage reporting
  • Set LOG_JSON=true for machine-parseable log format

Network Egress Summary

Mode Search Provider LLM Provider Outbound Calls
local-only Any Ollama (local) None
local-only Any Cloud (misconfigured) LLM only ⚠️
hybrid Builtin Ollama None
hybrid Tavily/Exa Cloud LLM Search + LLM
cloud-enabled Tavily/Exa Cloud LLM Full egress

Container Hardening

For production or sensitive environments:

# docker-compose.yml example
services:
  aideator:
    image: your-org/aideator:latest
    environment:
      APP_DEFAULT_MODE: local
      LOG_JSON: "true"
    # Block all outbound except localhost
    networks:
      - internal
    read_only: true
    tmpfs:
      - /tmp
    security_opt:
      - no-new-privileges:true

Recommendations:

  • Run with APP_DEFAULT_MODE=local in regulated environments
  • Use read-only filesystem with tmpfs for temporary files
  • Store secrets in a secret manager (Docker secrets, Vault, cloud KMS)
  • Keep external API keys empty for guaranteed offline operation
  • Apply network egress controls at the container/host level
  • Centralize logs in approved internal observability systems only
  • Review mode defaults before release (APP_DEFAULT_MODE and env files)

Reporting Vulnerabilities

Please report security issues by opening a private issue on the GitHub repository.